fix(scorers): preserve default LLM judge categories

[YusefSyed]

Summary

• let BaseScorer retain an LLM judge subclass's declared category when construction does not supply an override
• preserve both non-empty and explicit empty-string category overrides
• add focused regression coverage for all three constructor paths

Fixes #28.

Why

LLMJudgeScorer.__init__ previously defaulted category to "" and always passed that value to BaseScorer. Because BaseScorer uses None to mean “keep the class attribute,” direct construction replaced categories such as FactualityJudge.category == "MIT-3.1" with an empty string.

Changing the default to None uses the existing sentinel contract. Callers that explicitly pass a category—including category=""—retain the previous override behavior.

Validation

• uv run --extra dev pytest tests/test_llm_judge_categories.py -q — 3 passed
• uv run --extra dev pytest tests/test_groundedness_scorer.py -q — 6 passed
• uv run --extra dev pytest -q — 9 passed
• uv run python -m compileall -q rai_toolkit tests — passed
• uvx --from 'reuse[charset-normalizer]' reuse lint — REUSE 3.3 compliant

The existing test run reports 11 asyncio.iscoroutinefunction deprecation warnings from rai_toolkit/_tracing.py; this change does not introduce or modify them.

AI assistance disclosure

AI tools assisted with repository search, implementation drafting, and an independent fresh-context review. I directed the work, reproduced the bug on the pinned upstream revision, reviewed the source and final two-file diff, added the compatibility edge case raised during review, and ran the validation listed above.

[knisar]

This is the right fix and the right size. Using the existing None sentinel rather than adding a new one keeps category consistent with how BaseScorer already handles name, description and threshold, and keeping the explicit category="" override working means nobody's existing call site changes meaning. Three tests for three constructor paths is exactly the coverage this needed.

Verified locally: your 3 tests pass, the PR head suite is green at 9, and merging this into current main after #24 gives 18 passing. It merges cleanly without a rebase.

Worth stating for anyone reading this later, since it is a behavior change even though it is the repair: any judge constructed directly without an explicit category now reports its real MIT category instead of an empty string. So pipeline summaries key on MIT-3.1 rather than falling back to the class name, several direct judges in the same category now aggregate together, category-based policy triggers and composite weights start working, and trace names pick up the category suffix. No report field or type changes. Where same-category judges have different assessed-row counts, their now-correct co-aggregation can move an overall score and not just its label. That is what #28 was about.

Thanks for the clean writeup and for catching this. Squashing and merging. Fixes #28.